Hi! Welcome to the mobile version of Mosaic.

This project is an alternative version of Mosaic inspired by "The Building Blocks of Interpretability" from distill.pub.

The point of this visualization is to see what neurons in our network are highly activated for different images.

Depending on the parameters you've set, the image will show what neurons are most excited at that location in the image. For example, you might see images that look a lot like birds in the part of an image that contains birds on a birdfeeder:



You may also want to add other lower layers to see more abstract feature visualizations (or to see both at once!):



For a more complete explanation, please check out the distill.pub article.